Navigating Uncertainty: Gold Price Forecasting with Kolmogorov-Arnold Networks in Volatile Markets

This study presents a comprehensive evaluation of multiple forecasting models for predicting gold prices, including traditional MLP-based models like LSTM, N-HiTS, and TSMixer, as well as the Kolmogorov-Arnold Network (KAN), within the KAN-MLP Hybrid Multivariate Market Forecasting Framework. This framework integrates diverse data sources, including gold prices, external financial indices like the U.S. Dollar Index and silver prices, and sentiment analysis derived from textual data. After preprocessing, these features are fed into the models to assess their performance in handling multivariate time series data during periods of market volatility. Evaluation metrics-MAPE, RMSE, MAE, MSE, and MASE-were employed to assess the models’ accuracy and computational efficiency. Results indicate that KAN consistently outperforms other models, achieving the lowest forecasting errors across various feature combinations. For example, with only gold and the U.S. Dollar Index (USDI) as features, KAN achieved a MAPE of $\mathbf{4. 1 \%}(\pm 1.3)$ and RMSE of 4.0 (±1.2), highlighting its accuracy and stability. KAN also proved exceptionally efficient, with training times ranging from 0.5 to 0.8 seconds, far lower than N -HiTS and TSMixer, which required up to 63.6 seconds for training on more complex feature sets. In contrast, LSTM, N-HiTS, and TSMixer showed more variability in performance and required longer training times. Traditional ARIMA models performed poorly, with a MAPE as high as 68.2%, underscoring their limitations in handling complex, non-linear data. Despite KAN’s success, limitations exist, such as the exclusion of external macroeconomic and geopolitical factors and the need for further hyperparameter tuning. Overall, the KANMLP Hybrid Framework provides a powerful, flexible solution for financial forecasting, offering significant improvements in accuracy and computational efficiency compared to traditional methods.